Reflecting on the year 2023, the humanoid robot sector has seen a bustling scene of activity. Bolstered by policy support, breakthroughs in AI technology, and a wave of investments, numerous players have plunged into the fray, vying to up their game. Amidst this fervor, the industry has quickly ascended to become the focal point of industrial evolution.
From various perspectives, the rise of the robot industry to prominence wasn’t a stroke of luck but the result of years of maturation in technologies both within and outside the industry.
Image from Canva
Humanoid Robots Ride the AI Industry’s East Wind
As early as 1973, Ichiro Kato of Waseda University in Japan led his team to develop the world’s first full-scale humanoid intelligent robot—WABOT-1. Yet, half a century has passed, and the commercialization of robots remains a distant dream. The root issues? Cost and performance are still fundamental hurdles. Specific to robots, they encompass motion modules, sensory modules, and artificial intelligence modules as their key technological facets.
For traditional robots, the proficiency in one of these technologies often suffices for their operational value. For instance, industrial robots primarily focus on motion control technologies, while vacuum robots lean towards navigation sensing technologies. Compared to these, humanoid robots demand broader applicability, surpassing previous robots limited to specific scenarios and venturing into a wider array of applications. Consequently, the complexity of their technology has exponentially increased, necessitating robust data modeling and a profound understanding of language and commands. However, breakthroughs in large AI models have gradually provided innovative solutions to these once insurmountable problems.
From the earlier Transformer to GPT-4, as the model parameters have exponentially increased, large models have evolved from text to incorporate voice, vision, and other modalities, leading to a general-purpose AI direction. This integration enables humanoid robots to enhance their capabilities in voice, vision, decision-making, and control. Viewed broadly, this is but one aspect of AI aiding the rapid technological advancements in the robot industry, significantly reflecting in the physical embodiment of robots.
Initially, by emphasizing AI’s generalization abilities and mimicking human actions, humanoid robots have gained autonomous decision-making and self-learning capabilities, enhancing task integrity and continuity. Next, the focus is on end-effector efficiency. By prioritizing the precision of dexterous hand operations, under the computation and decisions made by the central processing “brain”, the actions of humanoid robots should be accurate, reducing error rates and increasing task completion accuracy. Lastly, the capability for perception-based motion control showcases the emphasis on all-terrain mobility akin to autonomous driving, where humanoid robots should navigate and control their movement based on environmental perception, hence improving mobility and task efficiency.
In essence, AI technology has gradually unraveled many challenges that traditionally hindered the advancement of humanoid robots.
New Challenges Emerge
From an industry perspective, the rapid evolution of foundational technologies has led stakeholders to revisit this familiar yet unfamiliar field. Yet, in terms of actual industrialization, humanoid robots still have a long way before becoming a mainstream household presence.
Initially, the crucial data for the intelligence of humanoid robots remains limited. ChatGPT’s rapid iterations were possible due to the vast amount of public domain data available on the internet for its direct harvesting, a luxury humanoid robots don’t share. Robots in reality are few, and even fewer are available for data collection, making it a significant challenge. Additionally, considering the barriers and “walled gardens” created by various robot manufacturers to protect their data, the fragmentation further complicates data acquisition, impacting iterative improvements across different robots.
Currently, a collective database from 34 global robot laboratories encompassing over 60 existing datasets — with more than a million data points collected from 22 different robots for over 150,000 tasks — stands as a beacon in robot model training. Google’s open-source robot training dataset, Open X-Embodiment, represents a flagship in this field. However, this dataset, primarily focused on routine operations, still has gaps in full-body coordination and walking balance among others.
Moreover, constrained by computing power, current humanoid robots cannot respond to commands in real-time. General-purpose humanoid robots need to achieve a control cycle of 500Hz, but Google’s RT-2 model in robot control cycles only reaches 3Hz, missing the mark by more than two orders of magnitude. Lastly, and most importantly, the cost factor, with prices often reaching several tens of thousand dollars, remains a significant barrier to widespread consumer adoption. This indicates that progress in humanoid robots is still quite limited.
Accelerating Domestic Replacement
In fact, as the domestic market heats up, the wave of accelerating domestic replacement in China’s humanoid robot industry grows stronger, pushing the button on speeding up localization.
Firstly, from the demand side, as domestic labor costs gradually rise, there’s a growing need for robots, fueling enthusiasm among enterprises to engage in robot development. Observing major industrialized countries, the rapid development phase of the robot industry was partly due to the rising labor costs since the 1990s. However, with major developing countries joining the WTO and releasing their demographic dividends, a significant global industrial boom ensued.
Now, as labor costs in countries like China gradually rise, the global focus shifts towards innovating in the robot industry as a trend. Objectively, industrial chain migration within China is inevitable, but on a global scale, there are few countries with a high-quality labor force and stable social environment like China’s. Recent years have seen some of China’s lower-end industries migrate to Southeast Asia with less-than-expected outcomes. Against this backdrop, expectations are pinned on suppliers remaining in China, many of whom have turned their focus to robotics, objectively propelling the development of China’s robot industry.
Secondly, from a technological and industrial perspective, the presence of numerous high-potential domestic suppliers and application scenarios is crucial for achieving domestic replacement in China. Looking at the specifics of humanoid robots, their core components like gearboxes, servos, and controllers account for 70% of an industrial robot’s cost, a figure likely higher given the added complexity of a humanoid’s joints and degrees of freedom.
In these areas, domestic suppliers have begun to emerge, with companies like Estun Automation, Harmony Drive, and Lead Intelligent having made significant strides, among others, in the gearbox field. Servo domain anticipations are high for Hiconics Drive Technology, boasting a substantial domestic market share. Although a dominant supplier in the controller domain has yet to emerge, the industry is not without alternatives. With a vast population and a developed manufacturing sector, China offers vast application scenarios for robots, presenting innate opportunities for the domestic replacement of humanoid robots.
In summary, propelled by various factors, the domestic replacement of humanoid robots is gaining momentum.
2024: The Spring of Humanoid Robots?
Amid rapid industry growth, there’s talk about 2024 being the foundational year for robots. However, considering the current state of affairs, 2024 appears more a year for the initial landing of humanoid robots rather than the dawn of their industrial era.
First off, the technology involved in humanoid robots is exceedingly complex, not solvable overnight. Industry insiders agree that the past year marked a period of rapid change, yet a genuine explosion is still afar, given no fundamental shifts have occurred within the industry. This is because humanoid robots encompass high-end manufacturing and artificial intelligence among many fields, where the breadth and depth of technology are extraordinary.
For instance, the recently public robot company UBTECH has been developing for over a decade since 2012, continuing substantial investments both in core technologies and industrialization. For the industry at this juncture, the priority lies in leveraging the massive attention to deeply refine technology and industrial iterations, directing more resources toward core technology research and overcoming crucial challenges, thus achieving a complete commercial loop from research and development to product, application, and service.
Secondly, the challenges relating to the embodied intelligence of humanoid robots, such as data and hardware costs, also require time to address. Regarding large models, as previously mentioned, suitable data across various intelligences remains scarce. Moreover, both cloud-based and edge computing involve significant computational power consumption. Solving general understanding across various scenarios also requires time.
Furthermore, enterprises need mature suppliers to iterate and control costs for drivers, gearboxes, joints, and dexterous hands, an area still under development.
In the realm of technology, the focus on hardware standardization and operational algorithm paradigms will likely become the core lever for industry-wide cost reduction in the foreseeable future. Though the blurry line between ideals and reality concerning robots has long since blurred, the distinction between the two remains, indicating an inevitably slow industrialization path for robots.